Richard Sutton Warns Synthetic Data Is a Big Mistake

Prominent computer scientist Richard Sutton, a pioneer of reinforcement learning and author of the influential 2019 essay "The Bitter Lesson," has criticized the AI industry's growing reliance on synthetic data. Speaking about his new venture, Oak Lab, which he co-founded with his former student Khurram Javeed, Sutton argued that using artificially generated data to train large language models is a fundamental error. He believes this approach cannot solve the data bottleneck facing AI labs because the physical world is vastly more complex than any simulation humans can create.

This perspective stems from the "Big World Hypothesis," a concept developed by Javeed and researchers in Alberta. According to Sutton, any simulated environment is microscopic compared to reality, failing to accurately represent physical friction, robotic motor behaviors, or the unpredictable minds of other people. Furthermore, Sutton points out a human bottleneck in the process. Training systems on synthetic data requires human experts to filter and evaluate the datasets, which prevents the technology from scaling efficiently. For instance, designing a drone that mimics a bat's echolocation or correcting self-driving simulations still relies heavily on human intervention.

Instead of relying on static, human-designed simulations, Sutton advocates for AI agents that learn continuously from their own direct experiences. He criticized current language models because their weights remain frozen after their initial training phase. To achieve true intelligence, Sutton argues that systems must adapt constantly without experiencing catastrophic forgetting, where new information erases old knowledge.

Sutton's team proposed a solution to this problem in the journal Nature, introducing a method they call "Continual Backprop." While he acknowledges that modern language models represent an impressive scientific breakthrough, he estimates they only account for "like 20% or a quarter of intelligence" because they lack the ability to learn dynamically in an ever-changing environment.